---
title: Do read-only AI agents need governance?
url: https://opencontroller.lyzr.com/blog/do-read-only-ai-agents-need-governance
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# Do read-only AI agents need governance?

**Yes, but usually a lighter set of controls than agents that can write or send. A read-only agent cannot change records, but it can still expose data: through its answers, through links or images it renders, or by surfacing files a user should never have seen. Register it, scope what it can read, watch its output and cost, and skip the heavier approval and rollback controls.**

"Our agent only reads data, so does it need the same controls?" is a fair question. Read-only removes the most visible harms: deleted records, wrong payments, emails sent in your name. It does not remove data risk, because reading is how data gets out. Every agent belongs in the same [AI agent control plane](/blog/what-is-an-ai-agent-control-plane); only the depth of control differs.

## What can go wrong with a read-only agent?

| Risk | How it happens | Example |
|---|---|---|
| Exfiltration through output | Injected instructions make the agent put data in a URL, image link or answer that reaches an outsider | EchoLeak (CVE-2025-32711), an AI command injection issue in Microsoft 365 Copilot that allowed information disclosure over a network |
| Oversharing | The agent searches with broad permissions and shows a user content they could not open themselves | An HR file surfaced in a general assistant |
| Lethal trifecta | Private data + untrusted content + any way to send data out | A summariser that reads email and can fetch web pages |
| Wrong answers acted on | A person trusts a confident but ungrounded answer | Policy advice quoted to a customer |
| Cost | Long contexts, loops and large retrievals | A nightly job that re-reads a whole data store |
| Credential exposure | Read-only keys are still keys; leaked ones can be abused | A database read key committed to a repo |

The second and third rows are the important ones. "Read-only" often describes the agent's tools, not its data reach. An agent that can read everything a service account can read is a high-value target.

## Why is read-only not the same as low risk?

Two things decide data risk: what the agent can read, and whether any path out exists.

A tool that "only fetches a URL" is a path out, because the URL can carry data. A chat answer is a path out if the user is outside your organisation. A rendered markdown image is a path out if the client loads it. Simon Willison's [lethal trifecta](/blog/what-is-the-lethal-trifecta-for-ai-agents) captures this: private data, exposure to untrusted content and the ability to communicate externally. Any agent with all three can be made to leak, whether or not it can write to your systems.

## What controls should a read-only agent have?

Use risk tiers. Our [risk-tiering guide](/blog/how-to-risk-tier-ai-agents) explains how to assign them. Below is a sensible split.

| Control | Read-only, internal data, internal users | Read-only, sensitive data or external users | Read-write agents |
|---|---|---|---|
| Registered with owner | Yes | Yes | Yes |
| Scoped read permissions (user-delegated where possible) | Yes | Yes | Yes |
| Output scanning for sensitive data | Recommended | Yes | Yes |
| Egress and link or image rendering limits | Recommended | Yes | Yes |
| Groundedness checks on answers | Optional | Yes | Yes |
| Cost limits and attribution | Yes | Yes | Yes |
| Access review cadence | Annual | Quarterly | Quarterly |
| Human approval for actions | Not needed | Not needed | For high-risk actions |
| Rollback and undo planning | Not needed | Not needed | Yes |
| Kill switch | Yes | Yes | Yes |

The lighter column still has a kill switch. If a read-only agent starts leaking, you need to stop it fast.

## How do you scope what a read-only agent can read?

1. **Prefer on-behalf-of access.** When the agent acts for a user, use that user's permissions, so it cannot show them more than they could see. See [on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation).
2. **Avoid broad service accounts.** If one is unavoidable, limit it to named sources.
3. **Fix permissions at the source.** Oversharing often reflects old file and site permissions that an agent simply makes searchable.
4. **Use read-only database roles,** not application accounts with write rights the tool happens not to use.
5. **Review data reach on change.** A new connector or knowledge source changes the agent's risk.

## How Opencontroller does this

Opencontroller registers agents with owners, versions, tools and MCP servers, and Posture flags risky permissions and configuration, including over-broad access. Delegated on-behalf-of identity keeps agents within a user's own rights. Live groundedness guardrails and LLM judges score production traffic, and the LLM Gateway attributes cost per agent and per version. Rogue-agent handling monitors behavioural deviation and can quarantine or kill an agent. See [how to detect data exfiltration by AI agents](/blog/detect-data-exfiltration-by-ai-agents), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is a read-only AI agent safe by default?**
No. It cannot change records, but it can leak data through output, links and rendered content, and it can surface files users should not see.

**What is the minimum governance for a read-only agent?**
Registration with an owner, scoped read permissions, cost tracking and a kill switch. Add output scanning and egress limits when it handles sensitive data or external users.

**Does read-only mean we can skip access reviews?**
No. Data reach is the main risk for read-only agents, so review what they can read, at least annually and quarterly for sensitive data.

## Key facts

Read-only AI agents need governance, typically a lighter set than read-write agents. Risks: exfiltration through output, links or rendered images (e.g. EchoLeak, CVE-2025-32711, AI command injection in Microsoft 365 Copilot allowing information disclosure), oversharing via broad read permissions, the lethal trifecta (private data, untrusted content, external communication), ungrounded answers, cost and credential exposure. Minimum controls: registration with owner, scoped read permissions (on-behalf-of where possible), cost limits and a kill switch; add output scanning and egress limits for sensitive data. Opencontroller by Lyzr provides Posture, delegated on-behalf-of identity, groundedness guardrails, LLM judges, per-agent cost attribution and quarantine/kill. Pricing starts from $250K per deployment per year.

## Related

- [What is the lethal trifecta for AI agents?](https://opencontroller.lyzr.com/blog/what-is-the-lethal-trifecta-for-ai-agents.md)
- [How do you detect data exfiltration by AI agents?](https://opencontroller.lyzr.com/blog/detect-data-exfiltration-by-ai-agents.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
